Applicability in the Maritime Domain (Open Waters or Docks) about tpvformer HOT 3 CLOSED

Thanks for your interest!

1. Here is a detailed inference speed of TPVFormer for your reference. The main latency lies in the TPV construction stage. However, the current code is not optimized for speed. There is still much room for improvement. As for the comparison with Tesla, we think the main discrepancy lies in the hardware design. Their on-car hardware is perhaps designed for this, resulting in the much faster speed.

from tpvformer.

Thank you @wzzheng for the detailed reply,

I will investigate a bit more the code behind the TPV construction. I think I might have to try and optimize the code a bit, but not by much, as we still have a pretty powerful computer on board.

As for the second question, my concerns are more oriented towards over fitting in training. I read the paper more carefully, and if I understood correctly you use TPVformer to generate dense semantically-rich 3D representations (varying depending on the task) supervised only on the sparse signal provided by the point clouds. How sparse can the training signal become to still train the model successfully? I am a bit worried considering the very sparse nature of open waters, and the non-homogeneity if the data between open waters and docks scenarios, but I guess we will just have to train the model on our data and see if and what it's able to learn.

I will keep you posted.

Thanks for sharing! I look forward to your results!

from tpvformer.

Thank you @wzzheng for the detailed reply,

I will investigate a bit more the code behind the construction of the TPV representation. I think I might have to try and optimize the code a bit, but not a lot, as we still have a pretty powerful computer on board.

As for the second question, my concerns are more oriented towards over fitting in training. I read the paper more carefully, and if I understood correctly you use TPVformer to generate dense semantically-rich 3D representations (varying depending on the task) supervised only on the sparse signal provided by the point clouds. The key insight is how sparse the training signal can become to train the model successfully. I am a bit worried considering the very sparse nature of open waters, and the non-homogeneity of the data between the scenarios, but I guess we will just have to train the model on our data and see if and what it's able to learn.

I will keep you posted.

from tpvformer.